On 10/30/2012 10:50 AM, Underwood, Keith D wrote:

...

for shared memory, we do agree that you need some sort of memory barrier within the GET operation for all local GETs. I agree that on x86-like architectures, this might not be needed.

Why is it in the GET instead of the sync that precedes it? Without the sync, the memory barrier in the GET does not have value, does it?

My point with the example I gave is that the synchronization might not be an RMA Win_sync. In my example, it was the hand waving. If we require the user to do MPI_WIN_SYNC before doing an MPI_GET from local windows, the shared memory problem will be resolved. However, as you pointed out, this still doesn't resolve the network problem unless the FLUSH guarantees that the data is available in remote memory and not stuck in the PCIe complex somewhere. -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji